Martin William

I’m building Inkling for the @MetaforDevs Global AI Developer Hackathon. 🧵 Don’t know what to study at uni? You’ve probably got an inkling. Swipe through cards, watch a cosy room fill with things you love, and discover degree subjects you didn’t know existed.

Why open weights? It's a tool for 16–18 year olds. Every text step runs on an open model, so a school or careers service could host it themselves and keep pupils' answers in-house. Next up: making it feel like a finished product.

Muse Image drew all 163 objects in one paper style. SAM 3.1 found and cut out each one. Now when you answer a question, the matching object drops into your room. Instantly, because it was all made in advance.

Muse Spark Open (Meta's open-weights model, on Baseten) wrote a card for every subject: what you actually study, who'd enjoy it, and which object goes in your room. Then it reviewed its own work, removed duplicates and rebalanced where things go.

Days 3 + 4 of building Inkling (missed a post, not the work) 🧵 The big idea: do the AI work once, so the game is free to play at any scale. Every UK degree subject now has its own paper object. 163 of them.

Wonder if there’s potential to go the decentralised federated route ? Many computers doing smaller chunks of computation and training ?

Day 3: only had a few minutes spare today so starting prototyping round 1 Tap your school subjects UI gets a good amount on screen at once with a box to add more

Yeah these are stunning

Tomorrow: the “your subjects” screen and the “big questions” round. Building Inkling for @MetaforDevs

Also gave every UK degree subject an interest profile, built from the jobs its graduates actually go into (HESA + O*NET). Lesson: art graduates’ first jobs are mostly sales and hospitality, so job data badly undersells how creative those subjects are. You still need the human signal.

Day 2/10 · Same room, 22 styles, one question: can a computer find every object in it? Muse Image drew a student halls room from an object sheet, then grew it by adding a guitar and a globe while changing only ~5% of the rest. SAM 3.1 checked every object was really there.

Next: an art style, the object sheet, and the first real room. Following along here every day until 4 Nov. Built with Muse Code, Spark, Wonder, Image and SAM 3.1. @cerebral_valley

Also loaded the UK’s official course data (HESA Discover Uni): 28,753 degrees at 315 providers. Lesson one: read the docs. The “satisfaction” column I nearly used was actually about students’ unions 🙃

Day 1: smoke-tested Muse Spark, Muse Image and SAM 3.1 through the Meta Model API. Image drew a fish tank. SAM found it again with 98% confidence. That loop (draw the room, then check every object is really there) is the heart of the build.

Why: in a Student Room survey, 1 in 5 students said they wouldn’t pick the same uni again, and most who regretted their course said the content wasn’t what they expected. Most course search tools assume you already know what to study. Inkling starts from what you’re into.

Had to submit an application and idea

🎉 just got approved for the @Meta Global AI Developer Hackathon This is going to be a fun one and looking forward to trying out their model API’s

Make sure you have @helpbnk account, go and sign up helpbnk.com/

You have 1 hour left to sign up worldsbiggestbusinesscompeti…

@simonsquibb is launching the Worlds Biggest Business Competition right now youtube.com/live/eI1wgyMGww4…